New safety-contract MARL framework cuts network security budget violations from 100% to 0.3%

A paper on arXiv (2606.13832) from Jose Luis Lima de Jesus Silva introduces a safety-contract graph multi-agent reinforcement learning (MARL) framework for autonomous network security response. The work addresses a critical flaw in reward-only MARL: while it improves security metrics, it consistently violates operational safety budgets—like Mean Time to Recover (MTTR), false-positive response limits, and firewall change-management disruption. The proposed framework separates simulator observations from reusable operational budgets using constrained optimization, graph attention network encoding, and counterfactual action screening. Two instantiations are evaluated: C-MAPPO-GAT, which adds Lagrangian cost control, and ACD^3-GAT (Adaptive Constrained Counterfactual Decisioning with Graph Attention Network), which further incorporates CVaR tail-risk estimation, opponent-belief state, and Graph Counterfactual Risk Propagation.

In CAGE Challenge 4 benchmarks across three 200-episode seeds, unconstrained methods (IPPO, MAPPO-GAT) violated the SOC downtime budget in 100% of episodes, with mean downtime proxy costs of 311-430 against a budget of 50. C-MAPPO-GAT slashed the violation rate to 0.3% and mean cost to 15.5—a 95.6% reduction. ACD^3-GAT achieved a 13.8% violation rate with mean cost of 48.2, positioning it on the safety-contract frontier rather than at the most conservative point. Stress tests with topology-seed and coupled adaptive Red-process attacks confirmed that safety-constrained policies degrade far less than reward-only ones. This work makes autonomous SOC response deployable by ensuring agents operate within real-world operational constraints.